you released a paper very recently uh called Polaris which essentially I won't steal your thunder but it stitches models together to ensure that safety accuracy low hallucination and so

on can you talk a little more about the Polaris architecture itself and what inherently within it makes sure that it covers the deficiencies of AI agents or LLMs otherwise

you know there's this notion that you can make one model that's safe but really if you look at all safety systems in all engineering environments the safest systems are

always redundant systems there's always multiple things helping to ensure the safety you know you're never like "Oh we just had one brake on the car so it doesn't run

away there you know on the on the roller coaster right?" No there's actually redundant brakes on the roller coaster there's redundant ways it can stop the thing safely there's

redundant seat belts there's a seat belt and a bar that holds you in like they double up on everything and we said how can we double up on everything

in healthcare how can we double up on all the risky tasks the critical care instruction tasks like we don't want to get that wrong we don't want to hallucinate

on that we don't want to be wrong on that and that's where Polaris was born and Polaris is a constellation of models where we run multiple models sometimes in serial

sometimes in parallel and then you get the answer and you check it and then another one checks it and then you do a voting and then you can test

against a known rubric and you can say "Okay are they all on track?" if any one of them's out of line you go "Wait something's wrong we better we

better hold this" and so it slows it down but makes it more safe and that's what Polaris is all about